Cattle identification is progressively shifting towards non-invasive biometric techniques. However, 2D face recognition for non-cooperative subjects like cattle is challenging due to variations in pose, illumination, and expression. To address these challenges, we present a pose-invariant face recognition approach that performs effectively even with limited data. The method begins with 2D feature extraction using robust matching techniques. Depth maps are then generated from 2D images using a pretrained model, enabling 3D keypoint matching. The resulting 2D and 3D matching scores are fused at the score level to improve identification robustness. We curated a dataset comprising 4,625 images from 50 subjects captured under diverse poses, environmental conditions, and lighting. The proposed model achieves Rank-1 and Rank-2 identification accuracies of 82.77% and 88.48%, respectively. Since it requires no training phase, the model is highly scalable and well-suited for growing datasets without retraining.